severian

Local-first long-term memory for Hermes agents: a canonical append-only store (SQLite or PostgreSQL+pgvector) with five-dimensional tenant/profile/collection/agent/session scope, hybrid BM25+vector retrieval with a re-ranker, memories that cite the observations they came from, per-turn user facts, real dates next to relative times, truth resolution when a newer fact replaces an older one, and a conflict note when two returned memories disagree. Fully offline by default; Qwen3-Embedding-0.6B semantic embeddings opt-in, locally or in a user-owned sidecar. On OmniMemEval's public long-term-memory leaderboard (its own models, prompts and full public sets) it scores 90.6 LongMemEval (1st), 87.3 LoCoMo (2nd) and 71.6 BEAM 100K (2nd, within noise of the lead); it re-ran LCM-X to check the board's records, and per-question results, a strict second marker and scripts are published in the repo. Disclosure — stores every conversation turn in one SQLite database under the Hermes root (`<hermes root>/severian`) shared by all profiles with rows scoped per profile, or in the PostgreSQL database named by `SEVERIAN_DSN`; the opt-in Qwen/Granite embeddings download a model from Hugging Face on first use, and the opt-in embedding sidecar sends memory text to the URL you configure.

Sahil-SS95Memoryv0.3.13Updated 12 days ago

About

Severian memory plugin (for Hermes Agent)

Drop-in memory provider for Hermes. Install from the plugin catalog:

hermes plugins install severian

Or from a checkout: copy this folder to $HERMES_HOME/plugins/severian/ and run hermes pm repair, which installs the
pinned severian package from PyPI into Hermes' managed environment.

Then set memory: provider: severian in config.yaml and restart Hermes.

See the repo root README for features, benchmark results and configuration.

Disclosures

Memory is local. Everything is stored on the machine: by default under the Hermes data root
(<HERMES_HOME>/severian/, SQLite), or in the PostgreSQL database you point SEVERIAN_DSN at. The engine makes no
third-party calls.
Recall injects context. The provider's purpose is returning saved memories and assembled context blocks into
agent turns. Nothing is sent anywhere beyond what the bullets below name.
Optional network paths, each one you switch on:
- SEVERIAN_EMBEDDING=qwen: downloads the Qwen3-Embedding-0.6B weights from huggingface.co on first use.
- SEVERIAN_EMBEDDING=remote-qwen: talks to your own sidecar socket (SEVERIAN_EMBED_SOCKET), no internet.
- SEVERIAN_USER_FACTS=1: sends conversation turns to the model backend Hermes is already configured with, in the
  background worker you run, to short facts.
- SEVERIAN_RERANK: loads a cross-encoder locally (weights download from huggingface.co on first use).
- severian serve: starts the HTTP API yourself; not started by the plugin.
Background worker is yours. The plugin never spawns hidden processes; the derivation worker runs when you or
your cron run severian work.
No telemetry, no self-updating, no credential files. Configuration is the SEVERIAN_* environment variables you
set. The one secret, SEVERIAN_DSN (optional; it can hold a database password), is read through Hermes's
profile-aware secret scope (agent.secret_scope.get_secret), never from config files the plugin reads.
No process-wide changes. The optional embedding loaders pass offline mode to the model loader rather than setting
environment variables for the Hermes process.